GLOBE cuts AI reasoning tokens 49.7% in benchmarks

GLOBE, a training-free early-exit algorithm for AI reasoning, cut output tokens by 49.7% across two models in an arXiv preprint published Oct. 20.

In the Oct. 20 arXiv preprint, GLOBE also reduced active end-to-end wall time by 51.0% and scored 80.58% across five tasks, which the authors described as comparable to Full CoT.

The method targets chain-of-thought prompting, where a model produces intermediate reasoning before the final answer. The authors said extra generated steps can raise serving costs and may move a model away from a correct interim result.

“GLOBE embeds each completed reasoning segment as a point on a unit sphere and reads the shape of the resulting semantic trajectory,” the authors wrote in the abstract. A unit sphere is a sphere with radius one.

The algorithm first generates short trial answers in a separate branch. If a trial answer passes a confidence threshold, the main response stops; if not, a trajectory phase looks for repeated reversals and minimal overall movement.

The five benchmarks included mathematical reasoning, tool calling and agent memory, according to the abstract. The arXiv page excerpt did not list the two model names or disclose any dollar cost comparison.

The finding is preliminary.

The authors are Xinyue Peng, Zhenlin Liang, Meng Cai and Hua Wang. The version listed as arXiv:2610.25970v1 was submitted Oct. 19 at 08:01:32 UTC.

Source: preprint, arXiv cs.AI, Oct. 20, 2026